Scalable Gaussian Processes for Modeling Heteroskedastic Spacecraft Data

Spacecraft sensor data frequently exhibit heteroskedastic noise, with measurement uncertainty varying across operating conditions, energy regimes, and signal strengths. Accurately modeling this variability is critical for reliable denoising, estimation, and downstream scientific inference. We present a scalable Gaussian process (GP) framework that extends the Vecchia approximation to accommodate input-dependent noise variance while retaining computational efficiency.

We demonstrate the approach using real spacecraft data from instruments aboard the International Space Station, where strong heteroskedasticity arises in low–signal-to-noise and high-energy regimes. Compared to homoskedastic GP models, the proposed method provides improved uncertainty calibration and more robust identification of physically meaningful signal. The framework scales to large datasets and is well suited for automated processing pipelines. Beyond spacecraft applications, the method is broadly applicable to remote sensing and other large, noisy scientific datasets with spatially or temporally varying measurement error.

Gaussian processes

Scalable inference

Heteroskedasticity

Uncertainty quantification

Variance surface estimation

Vecchia approximation